The K-GAM network is a neural architecture that realizes Kolmogorov's Superposition Theorem as a fixed, data-independent Koppen inner embedding (a space-filling-curve-like map from the multivariate input to a single coordinate) composed with a trainable univariate outer network built as a ReLU generalized additive model. Fixing the inner map makes the model low-parameter and its outer units parallelizable, giving an interpretable alternative to transformers; the same framework recasts a transformer as an embedding followed by kernel smoothing (Nadaraya-Watson attention).
Polson and Sokolov introduce the Kolmogorov GAM (K-GAM) network, a low-parameter, highly parallelizable alternative to transformers built directly from Kolmogorov's Superposition Theorem. The archite…